Memorization and Knowledge Injection in Gated LLMs
This addresses the issue of catastrophic forgetting in LLMs for applications requiring continuous learning from new experiences, though it is incremental as it builds on existing gating and low-rank weight techniques.
The paper tackles the problem of LLMs struggling to sequentially add new memories and integrate knowledge by introducing a continual learning framework, MEGa, which injects event memories directly into model weights using gated low-rank weights, and it outperforms baselines in mitigating catastrophic forgetting on fictional characters and Wikipedia events datasets.
Large Language Models (LLMs) currently struggle to sequentially add new memories and integrate new knowledge. These limitations contrast with the human ability to continuously learn from new experiences and acquire knowledge throughout life. Most existing approaches add memories either through large context windows or external memory buffers (e.g., Retrieval-Augmented Generation), and studies on knowledge injection rarely test scenarios resembling everyday life events. In this work, we introduce a continual learning framework, Memory Embedded in Gated LLMs (MEGa), which injects event memories directly into the weights of LLMs. Each memory is stored in a dedicated set of gated low-rank weights. During inference, a gating mechanism activates relevant memory weights by matching query embeddings to stored memory embeddings. This enables the model to both recall entire memories and answer related questions. On two datasets - fictional characters and Wikipedia events - MEGa outperforms baseline approaches in mitigating catastrophic forgetting. Our model draws inspiration from the complementary memory system of the human brain.